Regulation-based probabilistic substance quality index and automated geo-spatial modeling for water quality assessment

Artyom Nikitin, Polina Tregubova, Dmitrii Shadrin, С. А. Матвеев, Ivan Oseledets, Mariia PukalchikScientific 2021 · DOI

TL;DR

This paper introduces an automated, probabilistic framework for mapping water quality across large regions. By combining Multi-task Gaussian Process Regression (GPR) with a new Probabilistic Substance Quality Index (PSQI), the researchers can predict not just the concentration of 21 different water parameters, but the statistical likelihood that a specific location meets government safety regulations.

What Problem Does This Paper Address?

Traditional water quality assessment is a headache for environmental scientists. It usually relies on a "Water Quality Index" (WQI), which aggregates various chemical and biological measurements into a single score. However, standard WQIs have major flaws: they often rely on subjective "expert" weights, they struggle with parameters that have "optimal ranges" (like pH) rather than just "less is better" limits, and they rarely account for the inherent uncertainty in spatial sampling.

Furthermore, modeling how pollutants move across a territory is computationally expensive and often requires "handcrafting" mathematical kernels to fit the specific geography. The authors sought a way to automate this process while providing a more rigorous, regulation-aligned way to define "good" water.

Methods

The authors built a pipeline that moves from raw sampling data to a high-resolution (100m²) interactive map. The core technical pillars include:

  1. Multi-task Gaussian Process Regression (GPR): Unlike standard kriging (which only looks at spatial distance), this multi-task approach models the correlations between different chemical parameters. If Calcium and Magnesium levels are usually linked in a specific geology, the model uses that relationship to improve its predictions for both.
  2. Spectral Mixture Kernel: To avoid the tedious process of manually selecting a covariance function, the authors used a Spectral Mixture Kernel. This allows the model to automatically discover complex, non-stationary patterns in the data during training.
  3. Data Warping: GPR assumes a normal (Gaussian) distribution, but water data is often bounded (e.g., concentrations can't be negative, and pH is strictly 0–14). The authors applied a "warping" approach to map these bounded measurements into an unbounded space for the model, then transformed them back for the final results.
  4. The PSQI: This is the paper's most novel contribution. Instead of a simple weighted average, the PSQI calculates the integral of the model’s predicted probability density function over the "safe" range defined by Russian regulatory standards (SanPiN).

Datasets

The study utilized a dataset from the New Moscow region (Russia) collected between 2017 and 2018.

  • Samples: 1,526 water samples.
  • Locations: 448 unique geographic points (wells, rivers, and springs).
  • Parameters: 21 physicochemical properties, including Alkalinity, Hardness, pH, and various ions/metals (Na, K, Ca, Mg, Fe, Cu, etc.).

Key Insights and Ideas

  • Uncertainty as a Feature: Most maps just show a "predicted value." This framework provides a confidence metric. If the model predicts a location is "safe" but the variance is high (perhaps because there are no nearby sensors), the PSQI reflects that uncertainty.
  • Inter-parameter Synergy: The model successfully captured "side-effect" correlations. For example, it found high correlations (ρ > 0.8) between Chromium and Iron, and between Copper and Nickel. This suggests that pollutants often travel in "packets" based on their source (like industrial runoff).
  • Automation vs. Complexity: The authors found that increasing model complexity (more "mixtures" in the kernel) didn't always help. The best-performing model was relatively lean, using a single mixture component and a rank-10 covariance matrix, which prevented overfitting to noisy environmental data.

Key Results

  • Predictive Accuracy: The model achieved varying degrees of success across parameters, with a peak average $R^2$ score of 0.635 for Sulfate ($SO_4$). Some trace metals (like Chromium and Iron) were much harder to predict due to high noise and low variability.
  • High-Resolution Mapping: The team produced a spatial distribution map of the New Moscow region at 100m² resolution, comprising over 151,000 prediction points.
  • Validation: The PSQI showed a 0.68 Pearson correlation with the actual fraction of measured parameters that fell within admissible bounds, proving it is a reliable proxy for regulatory compliance.
  • Efficiency: Once trained, the system is fast. Evaluating the PSQI for the entire region took only 95 seconds, making it viable for real-time decision support systems.

Limitations and Future Work

The primary limitation is the "curse of dimensionality." As the number of water parameters increases, the mathematical integration required for the PSQI becomes exponentially more difficult. To solve this, the authors had to use a marginalization approach rather than a full joint probability calculation.

Additionally, the model struggled with "noisy" trace elements. The authors suggest that future work could incorporate more "process-based" data—like soil types, land-use maps, or elevation models—as additional inputs to the GPR to help the model understand why certain pollutants cluster in specific areas.

Summary updated

Abstract

Natural environments are recognized as complex heterogeneous structures thus requiring numerous multi-scale observations to yield a comprehensive description. To monitor the current state and identify negative impacts of human activity, fast and precise instruments are in urgent need. This work provides an automated approach to the assessment of spatial variability of water quality using guideline values on the example of 1526 water samples comprising 21 parameters at 448 unique locations across the New Moscow region (Russia). We apply multi-task Gaussian process regression (GPR) to model the measured water properties across the territory, considering not only the spatial but inter-parameter correlations. GPR is enhanced with a Spectral Mixture Kernel to facilitate a hyper-parameter selection and optimization. We use a 5-fold cross-validation scheme along with [Formula: see text]-score to validate the results and select the best model for simultaneous prediction of water properties across the area. Finally, we develop a novel Probabilistic Substance Quality Index (PSQI) that combines probabilistic model predictions with the regulatory standards on the example of the epidemiological rules and hygienic regulations established in Russia. Moreover, we provide an interactive map of experimental results at 100 m<sup>2</sup> resolution. The proposed approach contributes significantly to the development of flexible tools in environment quality monitoring, being scalable to different standard systems, number of observation points, and region of interest. It has a strong potential for adaption to environmental and policy changes and non-unified assessment conditions, and may be integrated into support-decision systems for the rapid estimation of water quality spatial distribution.
BibTeX
@inproceedings{nikitin2021regulation,
  author = {Artyom Nikitin and Polina Tregubova and Dmitrii Shadrin and С. А. Матвеев and Ivan Oseledets and Mariia Pukalchik},
  title = {{Regulation-based probabilistic substance quality index and automated geo-spatial modeling for water quality assessment}},
  booktitle = {Scientific},
  year = {2021},
  doi = {10.1038/s41598-021-02564-w},
}
Figure